Quaternion-based vision-transformer for polycrystalline EBSD scans pre-trained on large-scale synthetic data

Dimensionality reduction is crucial in materials science for extracting patterns from high-dimensional data. This is vital for optimizing material design spaces by cutting computational complexity and for integrating diverse data sources for advanced anomaly detection, identifying structural or functional material deviations. Traditional convolutional autoencoders focus on local features, failing to capture global contextual information essential for predicting material behavior under stress, where overall material properties, grain orientations, morphologies and sizes are key. To address this, a Quaternion-Based Vision Transformer Masked Autoencoder (ViT-MAE) is proposed for polycrystalline materials. This model processes quaternion-valued EBSD orientation maps with a 65% masking ratio, capturing both local and global microstructural features efficiently. Pre-trained on synthetic data using self-supervision, the model achieves robust generalization on real-world EBSD scans by minimizing a misorientation loss function on the quaternion manifold. This work lays the groundwork for advanced multimodal data analysis in materials science, offering an open dataset and pre-trained vision-transformer weights. The goal is to provide a reliable latent space for EBSD orientation maps, advancing future multimodal characterization through latent space merging. • A Quaternion-Based Vision Transformer Masked Autoencoder (ViT-MAE) was developed for polycrystalline EBSD scans. • Synthetic datasets are generated to pre-train the model. • This model captures both local and global microstructural features from EBSD orientation maps using a high mas- king ratio. • Validation on real-world EBSD scans demonstrated the model's generalization capabilities • The classification of out-of-distribution features is demonstrated in a zero-shot anomaly detection downstream task.